Upstream llama.cpp Activity

Active discussions

• Discussion #20969: TurboQuant early discussion
• Issue #20977: Feature request (Mar 25). mudler has experimental fork.

Contribution requirements (PR #19762, merged Mar 13)

To upstream a new quant type, must provide:

• GGUF conversion support
• Perplexity vs FP16 comparison
• KL divergence data
• CPU performance baselines

Relevant optimizations to rebase on

• PR #20962: Metal Tensor API — 26% mul_mat improvement
• PR #20609: MXFP flash attention SoA layout pattern

Key lesson from upstream

Custom quant types that aren't in Metal SET_ROWS whitelist silently fall back to CPU.
We hit this bug. Pattern confirmed by MXFP4 experience (PR #20609).